[PATCH] crypto: algboss - don't wait during notifier callback

From: Eric Biggers <redacted>

When a crypto template needs to be instantiated, CRYPTO_MSG_ALG_REQUEST
is sent to crypto_chain.  cryptomgr_schedule_probe() handles this by
starting a thread to instantiate the template, then waiting for this
thread to complete via crypto_larval::completion.

This can deadlock because instantiating the template may require loading
modules, and this (apparently depending on userspace) may need to wait
for the crc-t10dif module (lib/crc-t10dif.c) to be loaded.  But
crc-t10dif's module_init function uses crypto_register_notifier() and
therefore takes crypto_chain.rwsem for write.  That can't proceed until
the notifier callback has finished, as it holds this semaphore for read.

Fix this by removing the wait on crypto_larval::completion from within
cryptomgr_schedule_probe().  It's actually unnecessary because
crypto_alg_mod_lookup() calls crypto_larval_wait() itself after sending
CRYPTO_MSG_ALG_REQUEST.

This only actually became a problem in v4.20 due to commit b76377543b73
("crc-t10dif: Pick better transform if one becomes available"), but the
unnecessary wait was much older.

BugLink: https://bugzilla.kernel.org/show_bug.cgi?id=207159
Reported-by: Mike Gerow <redacted>
Fixes: 398710379f51 ("crypto: algapi - Move larval completion into algboss")
Cc: <redacted> # v3.6+
Cc: Martin K. Petersen <redacted>
Signed-off-by: Eric Biggers <redacted>
---
 crypto/algboss.c | 2 --
 1 file changed, 2 deletions(-)

diff --git a/crypto/algboss.c b/crypto/algboss.c
index 535f1f87e6c1..5ebccbd6b74e 100644
--- a/crypto/algboss.c
+++ b/crypto/algboss.c

@@ -178,8 +178,6 @@ static int cryptomgr_schedule_probe(struct crypto_larval *larval)
 	if (IS_ERR(thread))
 		goto err_put_larval;
 
-	wait_for_completion_interruptible(&larval->completion);
-
 	return NOTIFY_STOP;
 
 err_put_larval:

-- 
2.27.0.rc2.251.g90737beb825-goog

On Thu, Jun 04, 2020 at 11:52:53AM -0700, Eric Biggers wrote:

From: Eric Biggers <redacted>

When a crypto template needs to be instantiated, CRYPTO_MSG_ALG_REQUEST
is sent to crypto_chain.  cryptomgr_schedule_probe() handles this by
starting a thread to instantiate the template, then waiting for this
thread to complete via crypto_larval::completion.

This can deadlock because instantiating the template may require loading
modules, and this (apparently depending on userspace) may need to wait
for the crc-t10dif module (lib/crc-t10dif.c) to be loaded.  But
crc-t10dif's module_init function uses crypto_register_notifier() and
therefore takes crypto_chain.rwsem for write.  That can't proceed until
the notifier callback has finished, as it holds this semaphore for read.

Fix this by removing the wait on crypto_larval::completion from within
cryptomgr_schedule_probe().  It's actually unnecessary because
crypto_alg_mod_lookup() calls crypto_larval_wait() itself after sending
CRYPTO_MSG_ALG_REQUEST.

This only actually became a problem in v4.20 due to commit b76377543b73
("crc-t10dif: Pick better transform if one becomes available"), but the
unnecessary wait was much older.

BugLink: https://bugzilla.kernel.org/show_bug.cgi?id=207159
Reported-by: Mike Gerow <redacted>

I forgot about the person who originally filed the bugzilla bug.
Kai, if you're okay with it, let's also add:

Reported-by: Kai Lüke <redacted>

On Thu, Jun 04, 2020 at 11:52:53AM -0700, Eric Biggers wrote:

From: Eric Biggers <redacted>

When a crypto template needs to be instantiated, CRYPTO_MSG_ALG_REQUEST
is sent to crypto_chain.  cryptomgr_schedule_probe() handles this by
starting a thread to instantiate the template, then waiting for this
thread to complete via crypto_larval::completion.

This can deadlock because instantiating the template may require loading
modules, and this (apparently depending on userspace) may need to wait
for the crc-t10dif module (lib/crc-t10dif.c) to be loaded.  But
crc-t10dif's module_init function uses crypto_register_notifier() and
therefore takes crypto_chain.rwsem for write.  That can't proceed until
the notifier callback has finished, as it holds this semaphore for read.

Fix this by removing the wait on crypto_larval::completion from within
cryptomgr_schedule_probe().  It's actually unnecessary because
crypto_alg_mod_lookup() calls crypto_larval_wait() itself after sending
CRYPTO_MSG_ALG_REQUEST.

This only actually became a problem in v4.20 due to commit b76377543b73
("crc-t10dif: Pick better transform if one becomes available"), but the
unnecessary wait was much older.

BugLink: https://bugzilla.kernel.org/show_bug.cgi?id=207159
Reported-by: Mike Gerow <redacted>
Fixes: 398710379f51 ("crypto: algapi - Move larval completion into algboss")
Cc: <redacted> # v3.6+
Cc: Martin K. Petersen <redacted>
Signed-off-by: Eric Biggers <redacted>
---
 crypto/algboss.c | 2 --
 1 file changed, 2 deletions(-)

Patch applied.  Thanks.
-- 
Email: Herbert Xu [off-list ref]
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt